Stripe built an internal AI prototyping tool called 'Demos not memos' that lets product teams generate working UI prototypes from text prompts, replacing the written spec as the primary artifact of product definition. Owen Williams, who led the project, describes a workflow where a PM or designer describes a feature in plain language and the tool produces a clickable, realistic demo within minutes, using Stripe's actual design system components.
This matters because it collapses the gap between ideation and stakeholder alignment. Written specs get misread. Demos get built wrong because engineers interpret language differently than designers intended. A shared visual prototype, generated before a single line of production code is written, cuts revision cycles and surfaces disagreement earlier. The original conversation goes deep on how Stripe wired the tool into existing design and engineering workflows, which specific LLMs power the generation step, and how they handled the hard constraint of keeping outputs faithful to their component library rather than producing generic UI hallucinations.
The open question is how this scales beyond Stripe, a company with a mature, well-documented design system and engineering culture that already prizes internal tooling. Williams addresses this directly, and his answer about what prerequisites a team actually needs before this approach works is worth reading for any org considering a similar build. The full episode is on YouTube, Spotify, and Apple Podcasts.
[READ ORIGINAL →]